Machine Learning Meets Reality: Why Most Intraday Trading Strategies Die at the Broker’s Fee

For years, the promise of machine learning in finance has been seductive: feed an algorithm enough price data, let it detect hidden patterns, and watch it print money. A new study from researchers at Universidad Anáhuac México delivers a sobering reality check. Alan Gaona, Jennifer Rangel-Madariaga, and Lourdes Uribe systematically evaluated three machine learning models for intraday price direction forecasting across 40 financial assets, including stocks, foreign exchange pairs, and cryptocurrencies, and found that after incorporating realistic transaction costs, only a limited subset of strategies remained profitable. The findings, published in Frontiers in Applied Mathematics and Statistics, highlight a gap that many previous studies have overlooked: the difference between predicting the market and actually profiting from it.

The research team trained Decision Tree, Random Forest, and XGBoost models on two years of historical one-minute data spanning January 2022 to December 2023. Rather than estimating exact prices, the models were reframed as classification problems, predicting whether future price movements would exceed the asset’s transaction cost within forecasting horizons of 1, 3, 5, 10, 15, and 30 minutes. This approach provides a probabilistic estimate of trend direction rather than a point forecast, which has shown superior results in previous literature. The predictive variables were intentionally restricted to technical indicators derived from high-frequency price and volume data, including moving averages, VWAP, MACD, the Stochastic Oscillator, Bollinger Bands, and the Relative Strength Index, deliberately excluding fundamental variables such as earnings reports or macroeconomic indicators that update at substantially lower frequencies.

One of the most striking findings concerns the forecasting horizon. Across the majority of the 40 analyzed assets, the 30-minute horizon emerged as the best-performing option within the evaluated candidate set. Among the 24 stocks, 22 selected the 30-minute horizon, while all 10 foreign exchange pairs favored this configuration, and 4 of the 6 cryptocurrencies also preferred it. The researchers caution that because 30 minutes represents the upper boundary of the candidate set, this result should not be interpreted as establishing a globally optimal forecasting horizon. Rather, it indicates a tendency toward improved performance at longer horizons, suggesting that immediate price fluctuations at the one-minute scale contain substantial short-term noise that longer intervals can filter out.

The methodology employed a walk-forward rolling-window validation framework with Bayesian hyperparameter optimization, ensuring that all reported test predictions were strictly out of sample. Class imbalance, a common challenge in intraday data where small price movements are more frequent than extreme ones, was addressed through class weighting optimized jointly with the remaining hyperparameters. Notably, rather than optimizing standard accuracy metrics, the researchers conducted Bayesian optimization using the F-beta Score with beta equal to 0.10, a metric that prioritizes precision over recall. This design choice reflects the economic reality that the cost of false-positive trading signals is high, where a small number of adverse trades can offset a larger number of profitable ones.

Regarding model selection, XGBoost was the most frequently selected learning algorithm, chosen for 18 of the 40 analyzed assets, followed by Decision Tree for 13 assets and Random Forest for 9. However, the preferred model varied across asset classes. Random Forest was the most frequently selected model for cryptocurrencies, accounting for 3 of the 6 assets, while XGBoost dominated among foreign exchange pairs. This behavior suggests that no single learning algorithm is universally preferable for all financial instruments, reinforcing the importance of evaluating multiple model families when developing intraday forecasting strategies.

The predictive results revealed substantial heterogeneity across asset classes. Stocks exhibited the strongest overall predictive performance, with an average F-beta Score of 43.88 percent, concentrated primarily in the Information Technology sector. Cryptocurrencies achieved competitive predictive performance, particularly among major cryptocurrencies, with Bitcoin and Ethereum exhibiting precision values exceeding 50 percent. Foreign exchange pairs remained more difficult to predict, with exotic currency pairs showing markedly weaker results than major pairs. Yet the researchers emphasize that superior predictive performance did not systematically translate into economic profitability.

After incorporating realistic transaction costs, the economic picture became considerably more restrictive. Only 5 of the 40 analyzed assets generated a positive average return per trade, while 6 assets achieved positive cumulative profits over the two-year evaluation period. Among stocks, only Tesla, Apple, and Microsoft generated both positive average returns and positive cumulative profits, with Tesla exhibiting the strongest risk-adjusted performance as measured by the Sharpe and Calmar ratios. None of the analyzed currency pairs generated positive economic returns after transaction costs, providing a clear illustration of the distinction between predictive and economic performance. Cryptocurrencies exhibited the strongest economic results, with Bitcoin, Ethereum, and BNB producing the most favorable outcomes.

The researchers also introduced the Maximum Tolerable Slippage, a novel metric quantifying the maximum average adverse price movement per execution that each strategy can tolerate while preserving positive net profitability. Tesla exhibited the highest MTS among equities, indicating greater robustness to execution costs, while Bitcoin showed comparatively high tolerance among cryptocurrencies. Benchmark comparisons against buy-and-hold, naïve persistence-based, and random-direction strategies further demonstrated that the profitable machine learning strategies consistently outperformed random trading, providing evidence that the learned predictive signals contributed meaningfully to the observed economic performance.

The study’s conclusions carry important implications for both academic research and practical trading. The findings suggest that economically viable strategies emerge from the combined selection of an appropriate forecasting horizon, learning algorithm, and prediction target rather than from any individual model alone. The researchers note that predictive accuracy alone is insufficient to assess the practical value of a trading strategy, and that profitability should be evaluated jointly with financial risk, benchmark comparisons, and execution-related considerations. As machine learning continues to attract attention in quantitative finance, this work provides a comprehensive framework for jointly assessing predictive and economic performance under realistic market conditions, reminding us that the market’s efficiency, while perhaps not absolute, remains a formidable barrier to consistent abnormal profits.

Subject of Research: Computational Finance

Article Title: Intraday trading strategy selection under transaction costs with machine learning

Article References: Gaona, A., Rangel-Madariaga, J., & Uribe, L. (2026). Intraday trading strategy selection under transaction costs with machine learning. Frontiers in Applied Mathematics and Statistics, 12, Article 1888767. https://doi.org/10.3389/fams.2026.1888767

Image Credits: AI Generated

DOI: 10.3389/fams.2026.1888767

Keywords: machine learning, intraday trading, transaction costs, XGBoost, Random Forest, cryptocurrency, financial forecasting, Intraday, trading, strategy, selection, under